Health Insurance, Moral Hazard, and Managed Care 1

Transcription

1 Health Insurance, Moral Hazard, and Managed Care 1 Ching-to Albert Ma Department of Economics Boston University 270 Bay State Road Boston, MA ma@bu.edu Michael H. Riordan Graduate School of Business Columbia University 3022 Broadway, 608 Uris Hall New York, NY mhr21@columbia.edu June 1997; April The authors acknowledge helpful comments from the editor, two referees, David Bradford, David Dranove, Thomas McGuire, and seminar participants at Columbia, Harvard and Yale Universities, the Federal Reserve Bank of New York, Massachusetts Institute of Technology, Universitat Pompeu Fabra (Barcelona), Universitat Autonoma de Barcelona, and Universities of New Hampshire, North Carolina and Toulouse.

2 Abstract If an illness is not contractible, then even partially insured consumers demand treatment for it when the bene t is less than the cost, a condition known as moral hazard. Traditional health insurance, which controls moral hazard with copayments (demand-management), can result in either a de - cient or an excessive provision of treatment relative to ideal insurance. In particular, treatment for a low probability illness is de cient if illness per se has little e ect on the consumer s marginal utility of income and if the consumer s price elasticity of expected demand for treatment is large relative to the risk-spreading distortion when these are evaluated at a copayment that brings forth the ideal provision of treatment. Managed care, which controls moral hazard with physician incentives, can either increase or decrease treatment delivery relative to traditional insurance, depending on whether demand management results in de cient or excessive treatment. JEL Classi cation Numbers: I10 B80 Keywords: Insurance, Moral Hazard, Health Care, Managed Care

3 Before the proliferation of managed care, health insurance contracts relied heavily on deductibles, copayments, and other coverage limitations to control the provision of care. These demand management instruments cause consumers to limit their demand for health care, and thereby reduce premiums by restraining health care costs, but may expose consumers to large uninsured risks. In contrast, managed care plans give health care providers explicit nancial incentives, such as prospective payments, capitations, and cost sharing contracts, to limit expensive treatments. Clearly, these supply management instruments have become more important. Economic theory interprets the emergence of institutions as a response to market failures. In the health industry, market failure stems from the di culty of contracting on illness ex ante. 1 Due to this failure, ideal insurance, which shields consumers from the costs of e ciently-provided treatment, is infeasible. Second-best insurance pays only part of treatment expenses or limits coverage in other ways. But patients who do not fully face the cost of medical care may demand treatment ine ciently, in the sense that the cost of treatment sometimes exceeds the bene t. This problem is known as moral hazard. Demand and supply management policies in health care attempt to control moral hazard while partly insuring consumers against the risks of illness. 2 In the 1960 s, the health economics literature adopted the term moral hazard to describe the di culty of contracting over health status (Arrow, 1963, 1968; Pauly, 1968; Zeckhauser,1970), and argued that demand management at best only partly cures the moral hazard problem. The very rst lines of Zeckhauser s 1970 article crystallize what is now the conventional wisdom : Theprimarypurposeofmedicalinsuranceistospreadarisk,theriskofincurringsubstantial medical expenses. With risk-spreading, individuals will not pay the full amounts of such expenses. Insurance provision will thus introduce a perverse incentive toward overexpenditure if, as is usually the case, (1) the insured has substantial in uence over the amount that is spent on their own behalf in any particular medical circumstance, and (2) the level of reimbursement by the insurance plan is a positively associated function of the expenses incurred by its insured. Given that demand management does not solve the moral hazard problem, it is plausible to 1 See Grossman and Hart (1986) on the di culties of contracting on complex events. 2 The health economics literature has recognized that demand and supply management policies are complementary responses to market failure in the health industry. Ellis and McGuire (1993) emphasize the potential importance of nancial incentives on consumers as well as on providers. Ma and McGuire (1997) use this basic premise to investigate the interaction between optimal insurance and provider payment. Newhouse (1996) in a recent survey urges more study of supply side incentives. 1

4 interpret supply management (managed care) as an additional attempt to do so. On this interpretation, Zeckhauser s statement of the conventional wisdom suggests that moral hazard under demand management alone results in an excessive provision of treatment, and that managed care reduces this excess. We demonstrate that this is not necessarily so. Under some conditions, optimal demand management results in de cient treatment and managed care corrects this de ciency, leading to more treatment. We construct a model of insurance for a particular treatment, and characterize conditions under which optimal demand management under moral hazard results in an excessive or de cient provision of treatment. By excessive (de cient), we mean that a higher (lower) fraction of individuals receive the treatment compared to the ideal benchmark in the absence of moral hazard. We extend the model to allow supply management (managed care), and prove that this relaxes the moral hazard constraint and, under certain conditions, achieves ideal insurance. In this case, the introduction of managed care increases (decreases) the provision of treatment if treatment under optimal demand management is de cient (excessive), thus alleviating the market failure due to moral hazard. The provision of treatment can be de cient under optimal demand management if the illness per se has little e ect on the consumer s marginal utility of income. This possibility depends on the income e ect of the copayment: the patient s marginal utility of income is higher if the copayment is higher. 3 The intuition is roughly as follows. There are two important con icting e ects on consumer welfare of limiting treatment by charging a copayment higher than the one that elicits the ideal ( rst-best) provision. First, the consumer su ers directly more income risk, as indicated by a higher marginal utility of income after paying for treatment. Ceteris paribus, the cost to the consumer of this departure from optimal risk-spreading is greater, the greater is the income e ect of the copayment. Second, the consumer enjoys a lower premium because the insurance company has a lower expected cost. This second e ect is more pronounced the greater is the income e ect of the copayment relative to the marginal utility of income when an ill consumer declines treatment, because the consumer s demand response to the increased copayment is greater and translates into a larger premium reduction. Thus the premium reduction e ect is magni ed if illness per se does not very much increase the marginal utility of income. The optimal provision of treatment under 3 The health economics literature has paid little attention to the importance of income e ects on the demand for treatment. In fact, a standard text in health economics only brie y mentions income e ects (Phelps (1992), pp. 32, 301). There are some notable exceptions. De Meza (1983) considers intertemporal income e ects of savings and insurance, but assumes that illness severity is contractible so that optimal insurance involves an indemnity payment when the consumer is sick. Marshall (1976) also allows income e ects, focusing on how the optimal insurance gives the consumer an incentive to exercise care to reduce the likelihood of illness. 2

5 demand management is less than the ideal amount if the price elasticity of demand for treatment (and hence the premium reduction e ect) is large relative to the risk-spreading distortion. While our main results are developed for the case of an infrequent illness, for which these clearly are the most important e ects, we also show that the possibility of de cient treatment does not depend crucially on a small illness probability. The rest of the article is organized as follows. The next section lays out a model of illness, treatment, and insurance. Section 2 studies the rst best, when the treatment decision is fully contractible, and Section 3 characterizes the second best, when the provision of treatment is controlled indirectly by demand management. Section 4 compares the two, and establishes conditions under which the second-best provision of treatment with optimal demand management is more or less than the rst-best with ideal insurance. Section 5 employs these results as benchmarks to study supply management (managed care). Section 6 concludes by summarizing, and describing possible future research. An appendix contains the formal proofs of propositions. 1 The Basic Model We focus on optimal health insurance for a particular discrete treatment. 4 Implicitly, we hold xed other health and disability insurance covering the consumer and study insurance for this particular treatment in isolation. To be concrete and to focus our discussion, we interpret the treatment as being suitable for a speci c diagnosis. 5 For example, the treatment might be a new class of prescription drugs for high cholesterol (statins) described in a recent Wall Street Journal article. 6 We assume that a consumer becomes ill with probability, 0 < <1. For example, is the probability that a consumer is diagnosed as having a high cholesterol condition and is potentially a candidate for statins. Many of our results focus on the case where is small, which of course is often realistic. For example, according to the aforementioned Wall Street Journal article, 13 million 4 This contrasts with a more familiar reduced form approach in the health care literature, which assumes that a consumer s health is a function of undi erentiated health care expenditures. See, for example, Baumgardner (1991). Our discrete choice approach perhaps provides a more useful conceptual basis for empirical analyses of micro data which indicate whether a patient received a particular treatment or not. See, for example, Manning and Phelps (1979). Of course, it will be important in future theoretical work to study the bundled coverage of the panoply of treatments typical of health insurance policies. 5 More generally, the treatment in question might be suitable for a variety of diagnoses. For example, the treatment might be an extra day in the hospital due to complications from a surgical procedure. Our model is consistent with this broader interpretation. 6 See Price Prescription: Powerful Medications for Cholesterol Pose a Paradox for HMO s by Ron Winslow, Wall Street Journal, December 6, 1996, page A1. 3

6 people with serious heart problems are candidates for statins. While this is a lot of people, the probability that one of us will become a candidate for statins in the next year is hopefully quite small. It is easy to think of many other expensive treatments and tests that may be bene cial to an individual consumer with a small probability (organ transplants, neural surgeries, etc.). Our model focuses on health insurance coverage for one of these treatments in isolation. For simplicity, we assume that the cost of diagnosing the illness is zero. 7 Illness varies by its severity, which is described by a random variable ` with a cumulative distribution function F (`), and a density function f(`); with support on the positive real line. 8 As we make clear below, the severity of illness ` indexes the bene ts of treatment, and F (`) describes the distribution of bene ts across the population of patients. The consumer learns the bene ts of treatment (presumably in consultation with a physician) only after becoming ill. We assume that illness is not contractible, 9 and the insurance company can contract neither on the event of illness nor the realization of `. We adopt an eclectic approach to modeling the consequences of an untreated illness. Speci cally, consumers preferences are represented by a utility function of the form U(y a`) b`. Thevariabley represents the consumer s expenditure on other goods. The function U( ) is di erentiable, increasing and strictly concave, re ecting the consumer s risk-aversion to income uctuations, and a and b are non-negative parameters. Nested in this functional form are two special cases of interest. If b =0and a>0, then illness is completely equivalent to a loss of monetary income; we call this the monetary loss model. In this case, we sometimes set a =1as a normalization, measuring the loss of health resulting from illness in monetary units. The second special case is the utility loss model with a =0and b>0. Here, illness a ects an additive utility loss and does not directly e ect the consumer s marginal utility of income; we sometimes also use the normalization b =1, interpreting ` as an expected utility loss from forgoing treatment. This broad interpretation implicitly recognizes the uncertainties surrounding illness and alternative treatments. The general formulation allows for mixtures of monetary losses and (additive) utility losses from illness. The utility loss model seems more plausible than the monetary loss model even though both are special cases. The monetary loss model implies that the demand for treatment does not de- 7 A xed and known diagnostic cost will not change our qualitative results, as long as the outcome of the diagnosis is not contractible. 8 The random probability of illness e ectively puts a mass point at ` =0. It is easy to extend the model to allow for bounded severity, and also for a strictly positive lower bound of the support to be strictly positive, as for a Pareto distribution (Ma and Riordan, 1997). 9 If the event of illness were contractible, the optimal insurance contract could involve a monetary transfer payment to patients who become ill. The case is perhaps better understood as disability insurance. 4

7 pend on income. This seems unrealistic for many discretionary treatments. Manning and Phelps (1979) found signi cant income e ects for speci c dental treatments. The RAND Health Insurance Study generally found small but positive income elasticities, while subsequent studies using di erent methods and data found larger elasticities (Phelps 1992). In contrast, the utility loss model implies a positive income elasticity of the demand for treatment without restricting its magnitude. The utility loss model does imply that a consumer s attitude toward income risk is independent of illness per se. Obviously, this is an empirical question, but seems extremely di cult to answer. We have no apriorireason to think that the e ect of illness on risk aversion goes one way or another. Therefore, it seems reasonable, at least as a rst step, to ignore this e ect. For these reasons, we are partial to the utility loss model. We make several simplifying assumptions about treatment. First, the cost of treatment is xed at C>0. It will be useful to think of C as a large number, creating a signi cant demand for insurance, but not prohibitively large. 10 For example, the Wall Street Journal article says that statin therapy costs $700 a year for life. While this is an expensive therapy, some consumers bene ts obviously are much higher than $700 a year. Second, if the patient receives treatment, then the losses from illness are eliminated completely. 11 Thus, ` indexes the bene ts from receiving treatment. These bene ts vary across patients, implying that some patients are better candidates for treatment than others. 12 Third, we model the treatment decision as a binary choice. Either the patient receives treatment or not. 13 Consequently, for a xed copayment, there is a critical value of ` such that more severely ill patients demand treatment while less severely ill patients do not, as we show later. Finally, the delivery of treatment is contractible, meaning that the insurance company can prevent fraudulent claims for the reimbursement of non-existent treatment costs. 10 Obviously, this treatment cost cannot be too large. Otherwise, there would be little demand for insurance in our model. If C is su ciently small, then there is a demand for insurance to smooth the income uctuation resulting from illness and treatment. Interestingly, in the monetary loss model, the uctuation to be smoothed is the monetary equivalent of illness itself. In the utility loss model, the uctuation to be smoothed is the cost of treatment. 11 As mentioned above, we do not need to assume this in the utility loss model. In this special case, ` is the bene t of possibly imperfect treatment. 12 We emphasize that our interpretation of ` is the treatment s potential bene t to patients. One could interpret alternatively that ` refers to patients losses, although we do not think that this is always true. For some illnesses, it may be more natural to allow for multiple treatments when severity varies. See Chernew, Encinosa and Hirth, (2000). 13 A more general, mechanism design approach to optimal insurance would allow the insurance company to randomize the provision of treatment. We ignore this possibility because it is unrealistic. 5

8 2 Ideal insurance We begin by studying the ideal insurance contract in the absence of moral hazard. In this regime, illness and loss are assumed to be completely contractible; payments and treatment decisions can be contingent on the severity of illness. Under an ideal contract, the consumer pays a xed premium P, and receives treatment whenever the bene ts of treatment ` is above a xed threshold L. Moreover, when treatment is withheld, the consumer can be compensated by an indemnity payment, t(`), that depends on the severity of illness. In principle, the consumer receiving treatment can also receive a transfer payment. This possibility, however, plays no role in our model because treatment eliminates all illness losses. Therefore, under an ideal contract a consumer with income Y enjoys an expected utility given by the formula: ( Z ) L (1 )U(Y P )+ [U(Y P a` + t(`)) b`]f(`)d` +[1 F (L)]U(Y P ) : (1) 0 This expected utility formula is a weighted sum of two components. The rst component is the utility of a healthy consumer who has paid the insurance premium. 14 The second component (in curly brackets) is the conditional expected utility of an ill consumer, who receives treatment when the bene t is su ciently great, but otherwise is compensated only by the indemnity payment. The weights are the respective probabilities of illness and health. The insurance premium must cover the expected costs of treatment and indemnity payments. Thus, the premium constraint for the ideal contract is given by the inequality: " Z # L P t(`)f(`)d` +(1 F (L))C : (2) 0 The ideal contract maximizes expected utility subject to the premium constraint; it is the solution (P ;t (`);L ) to the problem of maximizing (1) subject to (2). The premium constraint is necessarily binding at an optimum. Proposition 1 The ideal insurance contract speci es (a) a treatment threshold that equates the marginal bene ts of treatment to the marginal cost, i.e. U 0 (Y P )al + bl = U 0 (Y P )C; (3) (b) indemnity payments that exactly compensate for the monetary losses from an untreated illness, i.e. t (`) =a`, and (c) a premium that exactly covers the expected costs. Finally, fewer patients receive treatment if the cost of treatment is higher, i.e. L is strictly increasing in C. 14 The variable Y can be interpreted as income net of premiums paid for other health insurance. Thus Y P is net income after the premium insuring for this particular treatment. 6

9 The characterization of the optimal treatment threshold needs a bit more interpretation. The marginal cost of treatment C is in monetary units. To measure the value of the cost of treatment C in utility units, it is multiplied by the marginal utility of income U 0 (Y P ). Equation (3) says that the utility of treatment evaluated at the threshold level of illness, [U 0 (Y P )a + b]l,equals the cost of treatment measured in utility, U 0 (Y P )C. In the monetary loss model (a =1;b=0), equation (3) states that treatment will be provided whenever the monetary loss from an untreated illness exceeds the monetary cost of treatment (` C). In the utility loss model (a =0;b=1), the treatment threshold is set equal to the cost of an untreated illness (L = U 0 (Y P )C), andthere is no indemnity payment for an untreated illness (t (`) =0). 3 Demand Management We now turn to optimal insurance when illness and loss are not contractible. The insurance company contracts with the consumer to reimburse part or all of the cost of treatment, but does not directly observe illness. The treatment decision itself is delegated to the patient, presumably in consultation with a physician. 15 We refer to this scheme of insurance as demand management. Under a demand-managed insurance contract, the patient pays a xed premium P up front, and copays D for treatment. Given this contract, the patient demands treatment when the bene ts exceed the utility cost of the copayment. Thus treatment is provided when the severity of illness exeeds a threshold value L satisfying the treatment constraint : U(Y P al) bl = U(Y P D): (4) This constraint states that a patient with illness severity L is indi erent about receiving treatment, and implicitly de nes L as a function of D and P. This function determines the expected demand for treatment by consumers who fall ill, which is equal to [1 F (L)]. As with ideal insurance, the premium must cover the expected cost of the insurance company. Since treatment is provided only when illness severity exceeds L,theex ante probability of treatment is [1 F (L)], and the insurance company breaks even if the premium covers its expected liability: 16 P = (1 F (L))(C D): (5) 15 Implicitly, we assume that the physician is a perfect agent for consumers. At the end of this section, we brie y discuss competition between self-interested physicians. In Section 5, we examine physician incentives under managed care. 16 The premium constraint, P (1 F (L))(C D), must be binding for an optimal insurance contract, because expected utility, given by (7) below, is decreasing in the premium. 7

10 The breakeven constraint de nes a tradeo between the premium and the copayment: 17 ½ dd = [1 F (L)] + f(l)(c ¾ dp dd The rst term in this expression is the direct e ect of shifting more cost to the patient, while the second is the marginal demand reduction from a higher copayment. Solving for the total e ect reveals that a higher copayment translates into a lower premium: 18 dp dd = [1 F (L)][aU 0 (Y P al)+b]+ f(l)(c D)U 0 (Y P D) [au 0 (Y P al)+b] f(l)(c D)[U 0 (Y P D) U 0 < 0: (6) (Y P al)] Consumer expected utility is a probability weighted average of utility when healthy and ill: ( Z ) L (1 )U(Y P )+ [U(Y P a`) b`]f(`)d` +[1 F (L)]U(Y P D) : (7) 0 The event of illness occasions two possible losses. First, a consumer su ers monetary (a >0) and additive utility (b > 0) losses from an untreated illness (when ` L). Second, the copayment reduces the disposable income of a patient receiving treatment (when ` L). Optimal demand-management sets a premium and copayment to maximize expected utility, given that consumers determine treatment and the insurance company breaks even. Thus, the optimal contract (P d ;D d ;L d ) maximizes (7) subject to (4) and (5). We call this the demand management problem. A rst result, traced to Zeckhauser (1970), is that consumers are only partially insured for the cost of treatment. The consumer faces some risk of untreated illness, and copays less than the full cost of treatment. The solution balances the negative consumer welfare e ect of a higher copayment against the positive e ect of a lower premium. Proposition 2 Optimal demand management partially insures consumers against the cost of treatment (0 <D d <C), and less severely ill patients decline treatment (L d > 0): The optimal copayment balances the expected utility cost of a marginally higher copayment against the corresponding bene ts of a lower premium, i.e. ( Z ) L dp (1 )U 0 (Y P )+ U 0 (Y P a`)f(`)d` + [1 F (L)]U 0 (Y P D) (8) dd where dp=dd is given by (6). 17 >From the = For a small illness probability, the e 0 = [1 F (L)]U 0 (Y P D) U 0 (Y P D) au 0 (Y P al)+b dp > 0 = U0 (Y P D) U 0 (Y P al) au 0 (Y P al)+b > 0. is of second-order importance. 18 dp=dd < 0 follows from the fact that the treatment constraint implies D al. 8

11 The proposition reveals two marginal e ects on consumer welfare of raising the copayment: a direct e ect of more income risk, and a premium reduction e ect. At an optimum, there is no need to weigh the marginal e ects on the demand for treatment, because these are of second-order importance. 19 The optimal copayment depends on the consumer s risk aversion and on the price elasticity of the expected demand for treatment, both of which depend on the curvature of the utility function. The well-known Arrow-Borch condition for optimal risk-spreading requires that marginal utilities of income are equal in all states of illness and health. However, a copayment will create a distortion from optimal risk-spreading across the states in which the consumer is healthy and the states in which the consumer is treated for illness. The magnitude of the Arrow-Borch distortion for a h i given copayment, U 0 (Y P D) U 0 (Y P ) 1, depends only on the curvature of the utility function on the domain between Y D and Y. The consumer s elasticity of expected demand for treatment with respect to the copayment (holding the premium constant), 20 f f(l) DU 0 (Y D) [1 F (L)] [au 0 (Y al)+b] g, also depends on the curvature of the utility function, as well as on the losses from untreated illness and on the severity distribution. These two magnitudes, the Arrow-Borch distortion and the price elasticity of demand, are prominent in determining the optimal cost-sharing ratio ( C D D ) when the probability of illness is small. Corollary 1 If the probability of illness is small, then the optimal ratio of cost shares for the insurance company and the patient is approximately equal to the ratio of the Arrow-Borch distortion and the price elasticity of demand for treatment, i.e. ( f(l d ) D d U 0 (Y D d )" # " ) C D d U 0 (Y D d # ) [1 F (L d )] [au 0 (Y al d )+b] D d = U 0 1 (9) (Y ) in the limit as! Taking the total derivative of (7) with respect to(d; P; L) yields ( Z ) L (1 )U 0 (Y P )+ U 0 (Y P a`)f(`)d` + [1 F (L)]U 0 (Y P D) dp 0 [1 F (L)]U 0 (Y P D)dD + [U(Y P al) bl U(Y P D)]f(L)dL: The treatment constraint (4) implies that the last term vanishes, leaving the two e ects identi ed in the proposition. 20 To derive this elasticity, di erentiate [1 F (L)] with respect to D, using (4) and the chain rule, and apply the de nition of elasticity. 21 The corollary is proved by substituting (6) into the rst-order condition in Proposition 2, dividing through by ; and taking the limit as! 0. In the limit, the premium is zero because the consumer never 9

12 The corollary illustrates neatly the con ict between providing insurance and controlling moral hazard. On the one hand, if the consumer is highly averse to income risk (a large Arrow-Borch distortion), then the insurance company should bear a high fraction of the treatment cost in order to better insure the patient. On the other hand, if the demand for treatment is sensitive to price (a high price elasticity of demand), then the consumer should face a substantial treatment expense in order to curtail an excessive demand. Optimal cost sharing balances these two concerns. We close this section on a technical note. The demand management problem is not in general a concave programming problem. Therefore, the rst-order condition for an optimal contract in Proposition 2 is not necessarily su cient, nor is the solution necessarily well-behaved. The following corollary provides an assumption under which this is not an issue. 22 Corollary 2 Assume that the following function is strictly increasing in D: ½ ¾ 1 D + U 0 (Y ) 1 U 0 [au 0 (Y al)+b] 1 F (L) ; (Y D) f(l) with L determined by the treatment constraint and P =0(the hazard rate assumption ). For su ciently small, the optimal threshold L d and copayment D d are unique, di erentiable and increasing in the treatment cost C. The hazard rate assumption requires that the inverse hazard rate, H(L) [1 F (L)]=f(L); does not decline too quickly. It is satis ed, for example, if H(L) is constant (exponential distribution) or increasing (e.g. Pareto distribution). becomes ill (P d =0). However, the premium is positive in the neighborhood of the limit. In general, the consequences of a change in the premium on expected consumer welfare are multifaceted, partly because the marginal utility of income varies across the di erent states of illness and health. However, some the premium e ects of a small variation of the insurance contract are more important than others, i.e. some are proportional to ( rst-order importance) while others are proportional to 2 (second-order importance). The small approximation focuses on the rst-order e ects. The second-order e ects that the approximation ignores can be described as follows: any change in the premium (P ) a ects consumer welfare di erently across the di erent states of illness and health because of the income e ects of the copayment (D) and untreated illness (al). Because the rst-order e ects are roughly proportional to and the second-order e ects are roughly proportional to 2, there is no particular reason to think that the second-order e ects dominate for plausible values of. 22 See Proposition 4 in Ma and Riordan (2000) for a formal proof. The assumption is not necessary for the result of the corollary. It is su cient that the demand management problem has a unique continuous global maximum. 10

13 4 Does Moral Hazard Increase Treatment? We now turn to our main question. How does moral hazard in uence treatment? Is the delivery of treatment under demand management more or less than that under the ideal contract? We make two points in this section. First, the provision of treatment can be either de cient (L d >L ) or excessive. Second, de cient treatment is a robust possibility, and occurs under various plausible conditions. We focus on the case of a small probability of illness. It is convenient to reformulate the demand management problem slightly. The treatment constraint (4) and the binding premium constraint (5) implicitly de ne the premium P and copayment D as functions of treatment threshold L. Solving for P and D as functions of L, and substituting these into the expression for consumer welfare (7) yields an expression for consumer welfare as a function of L alone, which we denote Z(L). Optimal demand management determines a treatment threshold that maximizes Z(L). Under the hazard rate assumption (of Corollary 2), Z(L) is a quasi-concave function and therefore achieves a unique local maximum at L d. We ask whether the derivative Z 0 (L ) is positive or negative. If Z(L) is quasi-concave and Z 0 (L ) > 0, thentreatment is provided de ciently under demand management (L d >L ), and conversely. Proposition 3 Under the hazard rate assumption, if the probability of illness is small, then treatment under optimal demand management is de cient if the cost-share ratio exceeds the ratio of the Arrow-Borch distortion to the price elasticity of demand, evaluated at the copayment for which the consumer demands treatment if and only if the illness severity (`) exceeds the ideal treatment threshold (L ), i.e. as! 0, L <L d if and only if ½ f(l ) D U 0 (Y D ) ¾ C D U0 (Y D ) [1 F (L )] [au 0 (Y al )+b] D > U 0 1 : (10) (Y ) where D satis es the treatment constraint (4) evaluated at P =0and L = L. The converse also holds. In the monetary loss model (b =0), the demand price for L is equal to the cost of treatment, i.e. D = C. In this case, inequality (10) obviously fails and we conclude that more treatment is provided under demand management than under the ideal contract. In this special case, an uninsured consumer seeks treatment e ciently, (i.e. only when the monetary bene t exceeds the cost a` > C) but this cannot be optimal. A risk averse consumer will always demand some insurance, resulting in an excessive provision of treatment; starting from D = C, the consumer is willing to pay a higher premium to achieve a lower copayment, which increases the demand for 11

14 treatment. 23 Corollary 3 In the monetary loss model (b =0), treatment is excessive under demand management (L d <L ). For the utility loss model (a =0), (3) implies that b = U 0 (Y )C=L as tends to 0, because the premium is proportional to. In this limiting case, inequality (10) is equivalent to C D C U 0 (Y D ) U 0 (Y D ) U 0 (Y ) > 1 F (L ) L f(l ) (11) with D satisfying U(Y ) U(Y D )=CU 0 (Y ). Substituting for C by the de nition of D,we can simplify (11) to U 1 0 (Y ) U(Y ) U(Y D ) D 1 U > 1 F (L ) 0 (Y ) L f(l ) ; U 0 (Y D ) whose left-hand side lies between 0 and 1 because U 0 (Y D )D >U(Y ) U(Y D ). The elasticity of demand with respect to the disutility ` is `f(`)=[1 F (`)]; we refer to this as the loss elasticity of demand. If demand is inelastic with respect to the utility loss, then the right-hand side of (11) is greater than 1, in which case (10) is violated and treatment is excessive. The other side of the coin is that moral hazard causes de cient treatment if the loss elasticity is su ciently large. Corollary 4 In the utility loss model (a =0and b =1), under the hazard rate assumption, if the illness probability is small, and demand is inelastic with respect to the additive utility loss evaluated at the ideal threshold ( L f(l ) 1 F (L ) < 1), then treatment is excessive under demand management (Ld < L ); if the loss elasticity is su ciently large, then treatment is de cient: The corollary can also be interpreted to say that treatment is de cient in the additive loss model if and only if the price elasticity of demand is su ciently high. The ordinary price elasticity of demand in the utility loss model ( L f(l ) [1 F (L )] D U 0 (Y D ) CU 0 (Y ) ) is proportional to the loss elasticity. Therefore, holding the utility function constant, an increase in the loss elasticity (which depends only on F ) translates directly into an increase in the price elasticity of demand. 23 The higher premium has a much smaller e ect on the demand for treatment because of the low probability of illness. Actually though, the following corollary does not require a small illness probability. For the monetary loss model, treatment is always excessive relative to the rst best. That is, the conditions in Proposition 3 are unnecessary. See Ma and Riordan (1997). 12

15 An interesting special case of the utility loss model occurs when the utility of income function exhibits constant absolute risk aversion, i.e. U(y) = exp( ry), where r>0 is the coe cient of absolute risk aversion. In this case, D =ln(1+rc), and (11) is equivalent to µ rc ln(1 + rc) 1+rC > 1 F (L ) rc rc L f( : (12) ) The right-hand side of (12) is the elasticity of demand with respect to the disutility of illness, evaluated at L. The left-hand side is an increasing, concave function of rc, ranging from 0:5 as r! 0 to 1:0 as r!1. Thus, holding the right-hand side constant, 24 we conclude from this special case that treatment is de cient if the consumer is su ciently risk averse and the loss elasticity is su ciently small. The intuitive explanation for this result is subtle. A higher degree of risk aversion clearly increases the magnitude of the Arrow-Borch distortion when D is raised above D to curtail treatment. However, a higher degree of risk aversion also corresponds to a greater income e ect of D on the demand for treatment. Thus, if r is higher, then a given increase in D curtails treatment more, resulting in a greater reduction in the premium. The latter e ect dominates, making the increase in D attractive to the consumer from an ex ante perspective. The above example and corollaries clearly demonstrate that de cient treatment is a robust possibility. This robustness argument can also be made in a more general way. Toward this end, for a given L, normalize the parameters a and b so that a C b = U 0 (Y ) L : (13) Thus the treatment threshold under the ideal contract is held constant as the parameters a and b vary according to this constraint. The inequality (10) can be expressed alternatively as C D n o 1 U 0 (Y ) > 1 F (L ) 1 f(l [au 0 (Y al )+U 0 (Y )f C ag]: (14) ) L U 0 (Y D ) The normalization (13) keeps L constant at the limit as tends to zero. An increase in a increases theright-handsideof(14)becauseu 0 (Y al ) >U 0 (Y ). Moreover, since P tends to zero with, the treatment constraint implies that D increases with a for the same reason. Therefore, the left-hand side of (10) decreases as the value of a increases, and (14) is more likely to hold when a is smaller. Now, xing values for C, U(Y ), U 0 (Y ), andd, we can make the left-hand side of (11) arbitrarily large by choosing a utility function with enough curvature on the interval [Y;Y D ] Note that L can be held constant as rc varies by adjusting Y suitably. 25 Meanwhile, the right-hand side can be held constant by adjusting a appropriately, and is constant if a =0. 13

16 This curvature is greater the more risk averse is the consumer on this interval. Putting together these implications of Proposition 3, we conclude that de cient treatment is a robust possibility over a range of values of a and b. Corollary 5 Under the hazard rate assumption and for a small illness probability, lesstreatment is delivered under demand management than under the ideal contract (L d >L ) if and only if the parameter a is su ciently small and the consumer is su ciently risk averse. The robust possibility of de cient treatment does not hinge on a small illness probability, as mentioned earlier. Let the utility function be logarithmic: U(Y )=ln(y), and assume a standard uniform distribution for losses: f(`) =1, F (`) =` with 0 ` 1 (thus dispensing with the hazard rate assumption as well). Let the cost of treatment be expressed as a fraction of income C = Y. Numerical results are displayed in Figure 1. The graphs describe the treatment thresholds L, as functions of : the lower one corresponds to the treatment threshold under the ideal contract; the higher one, demand management. Figure 1 shows that the threshold under demand management is higher for all values of. The graphs correspond to an expensive treatment: the value of was set at.65. Further numerical comparisons show that the di erence between the two thresholds (the second-best threshold subtracting the rst-best threshold) increases with. For small values of, this di erence is uniformly negative, while for high values of, it is uniformly positive. We have found that there are two local maxima for the numerical example and that the discontinuity in the second-best L represents a jump between them. Thus, the numerical example also illustrates that our de cient treatment possibility is robust even when the objective function is not quasi-concave. 26 In the introduction, we discussed the conventional wisdom that moral hazard causes an excessive provision of treatment. Of course, the conventional wisdom is imprecise, as are most conventional wisdoms. What is the relevant benchmark for determining if treatment is excessive under demand management? We have shown that the conventional wisdom is incorrect if the relevant benchmark is the delivery of treatment under an ideal contract when moral hazard is absent. 27 We have assumed implicitly that physicians act as perfect agents for the patients. In Ma and 26 It also illustrates how moral hazard can result in a rather extreme market failure no treatment or insurance. 27 Other benchmarks for the conventional wisdom are possible: (I) the amount of treatment that would be provided if consumers lacked insurance or (II) the amount of treatment that is e cient ex post, i.e. after consumers have paid the premium. The conventional wisdom according to interpretation (I) is obvious. The conventional wisdom according interpretation (II) is not necessarily correct. Once consumers have paid the insurance premium P d, it is socially e cient to provide treatment to an individual patient whenever the 14

17 Riordan (2000), we extend the utility loss model to include physician competition, and consider the other polar case in which the physicians are completely self-interested. In this alternative model, physicians compete for patients by adopting a practice style that commits them to a general level of service quality and a treatment threshold. Surprisingly, optimal demand management when pro t-maximizing physicians compete on practice style yields results almost identical to the case when physicians are perfect agents for consumers. Competition forces physicians to act in the interests of their patients, and in equilibrium, physicians adopt a treatment threshold that maximizes a consumer s ex post utility. 5 Supply Management and Managed Care A distinguishing feature of managed care is that health-care providers are given explicit incentives to limit treatments. Contracts with physician groups often feature a capitated payment for each patient but leave the physicians responsible for some or all of the treatment costs. This gives physicians an incentive to ration care in order to economize on treatment costs. Such supply management potentially complements and may even replace demand management as a method to control moral hazard. In this section, we introduce a pro t-conscious physician into the model and allow a role for managed care. For simplicity, we rely on the utility loss model. We argue that managed care bene ts consumers and may expand treatment. Indeed, managed care achieves ideal insurance in some cases. The technology of illness and treatment is the same as in our basic model. Consumer preferences have a utility loss representation: U(y) `. The insurance company contracts with the risk-averse consumer, and also with a risk-neutral physician (or physician group) to diagnose the patient and determine treatment. The physician learns the realization of ` from the diagnosis, and subsequently decides on treatment. We interpret managed care as a capitated payment arrangement in which the physician is severity of illness ` exceeds a threshhold L d satisfying U(Y P d ) U(Y P d al d ) bl d = U 0 (Y P d )C: If the probability of illness is small, then L d >L implies L d >L d. The reason is that, if the illness probability is small and P d is close to zero, then L d cannot be very much less than L. The formal proof is as follows. As! 0, P! 0 and P d! 0. Therefore, if L <L d in the limit, then a contradiction. bl d = U 0 (Y )C U(Y ) U(Y al d ) U 0 (Y ) C al d U 0 (Y )[C al ]=bl ; 15

18 credited a total payment of S + B for each diagnosed patient, but is liable for an amount B of the treatment cost. Equivalently, the physician is paid a xed fee, S, for a diagnosis, and an additional bonus, B, if the diagnosed patient does not receive treatment. 28 If the physician treats a patient with ` > L, then his expected payment is [S + F (L)B]. Without loss of generality, we assume that the treatment cost C is paid directly by the insurance company, i.e. the physician receives a net payment that deducts treatment cost liability from the capitation payments. The treatment decision is jointly determined by the physician and patient. To quote Arrow (1963, pp. 960): By certifying to the necessity of given treatment or the lack thereof, the physician acts as a controlling agent on behalf of the insurance companies. Needless to say, it is a far from perfect check; the physicians themselves are not under any control and it may be convenient for them or pleasing to their patients to prescribe more expensive medication, private nurses, more frequent treatments, and other marginal variations of care. In this spirit, we interpret the treatment decision as a collective decision that maximizes a weighted sum of the bene ts to the physician and the patient. 29 This might be interpreted as a reduced-form of a bargaining model, with the weights representing the bargaining strengths of the physician and the patient. Implicitly, we assume that bargaining occurs after the consumer becomes ill, but before the patient learns the severity of illness. Thus, the collective treatment decision in the utility loss model establishes a treatment threshold that maximizes the physician s expected treated payment minus the weighted patient s expected loss from illness, i.e. a value of L that maximizes ( Z ) L S + F (L)B µ `f(`)d` +[1 F (L)][U(Y P ) U(Y P D)] : 0 Themagnitudeoftheweight µ captures the bargaining strength of the patient, i.e. the higher the value of µ the greater the weight given to the patient s expected loss. The solution to the collective decision problem balances the physician s bonus payment against the treatment bene ts of the marginal patient: B µ [L U(Y P )+U(Y P D)] = 0: (15) 28 In principle, the contract could specify a payment to the physician even if the consumer does not become ill. However, under the limited liability assumption introduced below, this payment must be zero. 29 Ellis and McGuire (1990) use a similar assumption. 16

19 This optimality condition is analogous to a treatment constraint in the demand management problem. If B =0, then it is identical. If B>0, so that the physician is at least partially liable for the treatment cost, then it is weaker (L >U(Y P ) U(Y P D)) and some patients are rationed, i.e. a patient who wants treatment does not get it. 30 The other constraints on contractual relationships of the insurance company are as follows. First, the premium and expected copayment must cover the expected costs to the insurance company: P = fs + F (L)B +[1 F (L)] (C D)g: (16) Second, the physician must nd the relationship pro table: [S + F (L)B] U: (17) The term U in this individual rationality constraint is the opportunity cost of the physician s time and e ort in diagnosing and treating the patient. Finally, nancial constraints put lower bounds on payments to the physician, i.e. S S S + B S: (18) The second of these limited liability constraints is generally slack, and will be ignored. The rst may or may not be slack. The assumption that S is bounded from below by S allows the interpretation that the physician must be guaranteed a minimum compensation for the opportunity cost of diagnosis. 31 The optimal insurance contract maximizes the ex ante expected utility of the patient, given by (7) after setting a =0and b =1to conform with the utility loss model, subject to the treatment, individual rationality and limited liability constraints, given by (15) to (18). The problem can be simpli ed further. At an optimum, the premium constraint (16) must bind. Using this and the treatment constraint to eliminate S and B from (17) and (18) leaves a modi ed individual 30 The physician has no incentive to treat the patient and then do not report the treatment to the insurer in order to collect the bonus. This is because the physician would have to be responsible for the treatment cost. 31 Alternatively, the limited liability constraint can be interpreted as capturing risk aversion in a crude way; that is, the physician is risk neutral with respect to income variation above some critical level, and extremely risk averse for income variation below this level (Sappington, 1983). What is crucial for our analysis of managed care is that the physician is less risk averse than the patient with respect to the uncertain cost of treatment, and therefore, better able to absorb this risk. This is natural because the physician can diversify the treatment cost risks across a population of patients, and physicians can further pool this risk within a group practice. 17

20 rationality constraint, P U + [1 F (L)](C D); (19) and a modi ed limited liability constraint P fs + µf(l)[l U(Y P )+U(Y P D)]+[1 F (L)] (C D)g : (20) This reduces the managed care problem to choosing P, D, andl to maximize (7) subject to the modi ed individual rationality constraint (19) and the modi ed limited liability constraint (20). The managed care problem has the same structure as the demand management problem analyzed in Sections 3 and 4 (with S and U normalized to 0), except that the modi ed limited liability constraint (20) relaxes the treatment constraint (4). Therefore, since (4) was binding in the demand management problem, managed care must improve ex ante consumer welfare. We expect this relaxation of the treatment constraint to move the optimal treatment cuto in the direction of ideal insurance. Therefore, if optimal demand management results in de cient treatment relative to the ideal contract (e.g. as in Corollary 4), managed care should cause an expansion of treatment. Our next result shows that ideal insurance sometimes can be achieved when supply management is possible. Consider a relaxed managed care insurance problem in which the modi ed limited liability constraint (20) is ignored. That is, consider the maximization of (7) subject only to (19). The only di erence between this problem and the maximization problem for the ideal contract in Section 2 is that the premium must also compensate for the physician s opportunity cost of diagnosis. Proposition 1 applies directly, and the solution to the relaxed program speci es the ideal treatment threshold, L ; and a premium, P,thatsolveL = U 0 (Y P )C and P = U + [1 F (L )]C. The ideal copayment is zero. 32 Proposition 4 The optimal treatment decision, premium, and copayment under managed care are the same as for the ideal contract if and only if the minimum payment for diagnosis is low enough so that the limited liability constraint does not bind, i.e. [S + µf(l )L ] U. Under what conditions is the the limited liability constraint slack and ideal insurance possible? If copayment is set at zero, then the value of B must be set at µl to satisfy the treatment constraint (15) for the implementation of L. For the physician s reservation wage constraint to hold, the value of S must be set to U F (L )B = U µl F (L ). If this value of S is at least S, then the 32 The formal proof is straightforward and omitted. 18